Decoding preparatory movement state-based motor imagery with multi layer energy decoder Authors: Yuxin Zhang, Mengfan Li, Miaomiao Guo, Guizhi Xu, Alan Wang Published: 2025 ABSTRACT BackgroundMotor imagery (MI) is a widely used paradigm in brain-computer interface (BCI) research due to its potential applications in areas such as motor rehabilitation. As a purely cognitive process, MI produces low-amplitude, non-stationary EEG. Despite improving accuracies, cross-subject variability and limited generalization continue to motivate approaches that strengthen MI representations and enhance system robustness.MethodsWe designed a task-guided preparatory movement state-based motor imagery (PMS-MI) paradigm that elicits a brief motor preparatory state before MI and captures EEG features from both the preparation and imagery phases. To decode the features effectively, we introduced a multilayer energy decoder (MLED) that integrates graph signal processing (GSP): EEG is modeled as intra- and cross-frequency multilayer brain networks, and a graph Fourier transform (GFT) projects the signals into network energy features before classification. We benchmarked the PMS-MI paradigm and the MLED method across multiple time window lengths using a panel of classical and deep-learning classifiers.ResultsThe PMS-MI paradigm elicited significant energy variations during the movement preparation phase and induced earlier event-related desynchronization (ERD) with broader frequency band activation during MI, compared to traditional MI paradigms. Classification performance using CSP in the PMS-MI paradigm surpassed that of the traditional paradigm at all time windows. Further accuracy improvements were achieved with the MLED method. Brain network analysis revealed distinct neural representations between the preparation and MI phases, and MLED effectively captured these differences. Feature fusion of preparation and MI stages resulted in classification accuracies exceeding 85% for both 1 s and 4 s windows. The results demonstrate that both algorithmic design and paradigm choice play important roles in MI EEG decoding, with their relative contributions varying across temporal windows and experimental conditions.ConclusionsIntegration of preparatory movement states into the movement imagery process can generate distinguishable features at different stages and improve the classification performance of BCI systems. The proposed PMS-MI paradigm, combined with the MLED decoding method, provides a promising direction for developing more accurate and robust BCIs, particularly in the context of neurorehabilitation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-025-01837-z. Background Brain-computer interfaces (BCIs) enable communication and control between the human brain and external devices by decoding brain activity into machine commands, bypassing physical interaction [1, 2]. Electroencephalography (EEG) signals are widely used in BCIs research due to their noninvasiveness, low cost, portability, and high temporal resolution. In recent decades, BCIs have found successful applications in medicine as tools for rehabilitation and communication in individuals with motor impairments [3, 4]. Beyond medicine, BCIs are also used in fields such as entertainment [5], virtual reality [6], and education [7]. Among the various BCI applications, motor-BCI for limb rehabilitation have received the most research attention. By detecting movement intentions, MI-BCI can be integrated with FES neuroprostheses (hybrid neuroprostheses) or therapeutic exoskeletons within therapist-supervised rehabilitation to support recovery and improve motor capacity [8, 9]. Event-related desynchronization/resynchronization (ERD/ERS) in the α and β frequency bands [10] and low-frequency motor-related cortical potentials (MRCPs) in the sensorimotor cortex are key neural markers of motor execution (ME) and motor imagery (MI) [11]. Motor-BCI decode users' movement intentions by capturing these neural markers, thereby enabling control of external devices. Effective encoding and decoding of these signals is essential for accurately detecting movement intentions. However, motor-related signals represent endogenous neural activity, which is inherently weaker compared to evoked responses. Additionally, most motor-BCI users are patients unable to perform actual physical movements and thus can only generate movement intention signals through MI. MI-based mental activity, however, is challenging to control and elicit reliably, with some individuals experiencing BCI illiteracy or difficulties in manipulating the system effectively [12]. This makes the efficient use of MI-based BCI systems particularly demanding. To address these issues, many researchers have focused on developing MI paradigms (to enhance neural encoding) and refining decoding algorithms. The former aims to enhance neural representations, while the latter seeks to improve the recognition accuracy of movement intentions. Considerable research has been conducted on MI paradigm design to facilitate mental imagery and elicit more robust brain signals. For instance, Yi et al. designed a compound limb MI paradigm that generated more pronounced EEG oscillatory patterns than simple limb MI [13]. Qiu et al. developed a paradigm enabling users to imagine hand movements following the stroke order of Chinese characters, which facilitated the regulation of brain activity [14]. Yi et al. also proposed a novel synchronized hybrid BCI that integrates electrical stimulation into an MI-BCI system, demonstrating that combining MI with electrical stimulation significantly outperformed conventional MI-BCI approaches [15]. Bian et al. introduced a dynamic visual-aided paradigm with complex imagery tasks, eliciting stronger cortical activity in core motor-related regions [16]. In neurophysiological research, Wright et al. found that combining action observation (AO) with MI produced more significant cortical motor activity than either AO or MI alone [17]. In our previous study, we demonstrated that a preparatory movement state can enhance premovement EEG representations. We designed a visually guided motor task that incorporated the preparatory state into a premovement BCI paradigm, showing that, compared to spontaneous premovement, prepared premovement elicited more pronounced low-frequency MRCPs and high-frequency ERD features [18]. Integrating the preparatory movement state into the premovement encoding paradigm effectively enhances neural representations, ultimately improving motor-BCI performance. However, while the preparatory movement state can enhance premovement EEG representations, there remains a gap in understanding its effect on motor imagery. Further investigation is needed to explore its impact on neural representations and classification outcomes. To aid readers, we provide an overview of MI paradigms and classification algorithms in the Supplementary Material (Table S1 and S2). As mentioned before, to enhance the efficient use of MI-BCI systems, numerous studies have investigated decoding algorithms in addition to paradigm design. Given that the ERD/ERS phenomenon occurs during MI tasks, resulting in energy fluctuations across different brain regions, various feature extraction methods have been employed to capture and classify these energy changes. Common spatial pattern (CSP) is the most widely used spatial filtering technique for improving the signal-to-noise ratio of EEG power signals [19]. CSP operates by estimating pairs of projection vectors (also called spatial filters) from multi-channel EEG signals across two classes (e.g., two distinct MI tasks). Typically, CSP is applied to multichannel EEG segments that have been bandpass filtered within a relatively broad frequency range (e.g., 4–40 Hz). However, the effectiveness of CSP heavily depends on the chosen frequency band. Since the optimal frequency band is generally subject-specific, a fixed, wide frequency range cannot be universally determined for all cases [20]. Consequently, classification accuracy may be compromised if the selected filter band fails to encompass all discriminative features or includes redundant information. To address this issue, extensive research has focused on optimizing the filter band to extract more MI-relevant EEG features when using CSP [21–23]. For instance, Ang et al. developed filter bank CSP (FBCSP), which optimizes the filter band by leveraging the correlation between CSP features extracted from different frequency bands [24]. Additionally, as some channels contain redundant information and noise, channel selection methods based on classification algorithms have been proposed to further improve the classification accuracy of MI-BCI systems [25–27]. Most EEG-based MI decoding studies have primarily focused on channel-level analysis, wherein features are extracted from each EEG channel individually, with researchers mainly concerned about analyses within individual frequency bands. However, these features predominantly emphasize local brain activation and concentrate on analyzing the topology of single-layer brain networks, overlooking the fact that the brain operates as a complex non-linear system [28, 29]. The dynamic variations of MI may not be adequately captured by features derived from a single modality. Cross-frequency coupling of EEG signals reflects the synchronization of neuronal oscillations [30], which is crucial for the transmission and integration of information across different brain regions. Thus, both intra-frequency and cross-frequency coupling of EEG signals can aid in comprehensively exploring the brain's functional networks and identifying discriminative feature signatures. For amplitude envelope correlation (AEC), Hipp et al. demonstrated that band-limited power envelopes exhibit frequency-specific large-scale coupling in human MEG, establishing AEC as a robust functional-connectivity measure [31]. For amplitude–amplitude coupling (AAC), Knyazev et al. showed that source-level cross-frequency AAC in EEG tracks individual differences and clinical risk factors for affective disorders, illustrating the broader utility of AAC beyond MI-BCI [32]. Tewarie et al. [33] proposed a multilayer networks framework, emphasizing the significance of the multilayer network model by demonstrating that there is a strong dependency between intra-layer structure and inter-layer coupling. This finding suggests that networks obtained from different frequency bands are not independent entities. Although multilayer networks encode rich information in matrix representations, they often fail to efficiently extract actionable features. Graph signal processing (GSP), an emerging field, provides tools for analyzing signals in irregular domains, such as EEG and functional magnetic resonance imaging (fMRI). By combining graph theory and signal processing, GSP offers a unified framework [34]. A key component of GSP is the Graph Fourier Transform (GFT), a novel mathematical approach that extends traditional transforms to accommodate signals defined on graph structures. Beyond fundamental theory, existing research has applied GSP to decoding MI based on EEG. For instance, Feng et al. combined GFT with cross-band coupling features to decode voluntary and involuntary upper limb MI [35]. Annaby et al. modeled multi-channel EEG as directed/undirected graphs, extracted directed graph Fourier spectral features, and employed extreme learning machines for MI classification, achieving superior performance compared to classical baseline models [36]. Given that MI leads to variations in brain energy, the established multilayer network can be leveraged to derive the associated GFT basis functions, ultimately facilitating the extraction of intricate energy signals. Building on the challenges and findings related to MI-BCI encoding and decoding discussed in the previous sections, this study proposes a preparatory movement state-based motor imagery (PMS-MI) paradigm. We investigate the neural representations associated with both the preparation and MI processes within this framework. Furthermore, we introduce a new decoding model, the multi layer energy decoder (MLED), designed to extract features from EEG signals, and validate its performance using both the classical paradigm and the proposed paradigm. To the best of our knowledge, this represents the first exploration of a novel paradigm and feature extraction method for MI. Relationship to our prior work: our previous study focused on premovement intention detection in a button-press task and showed that a visually guided prepared state enhances MRCPs/ERD representations and yields a 4.67% accuracy gain over spontaneous premovement. In contrast, the present work targets a different stage and task: we integrate a PMS-MI of elbow flexion and evaluate it against a classical AO-MI condition. Methodologically, beyond conventional CSP, filter-bank, Riemannian, and deep-learning pipelines used previously, we introduce a MLED that explicitly models intra- and cross-frequency brain networks and performs GFT–based energy projection before classification. This stage-specific (preparation and MI) decoding and network-energy formulation differentiate the current study from [18] in both paradigm and decoder design. The main contributions of this paper are as follows:Inspired by our previous work, we developed a paradigm that integrates the preparatory movement state into MI and compared it with the traditional paradigm. We investigated the neural representation of the preparation state and its influence on the MI process.We propose a novel decoding framework that combines intra-frequency and cross-frequency brain networks with GSP to extract complex energy fluctuations from the brain. This framework employs spatial filters to enhance spatial differences between tasks while capturing both local and global energy features.We show that movement information is present during preparation, and that merging the preparation features with the MI features after separate processing leads to better classification for a range of MI window lengths. Inspired by our previous work, we developed a paradigm that integrates the preparatory movement state into MI and compared it with the traditional paradigm. We investigated the neural representation of the preparation state and its influence on the MI process. Inspired by our previous work, we developed a paradigm that integrates the preparatory movement state into MI and compared it with the traditional paradigm. We investigated the neural representation of the preparation state and its influence on the MI process. We propose a novel decoding framework that combines intra-frequency and cross-frequency brain networks with GSP to extract complex energy fluctuations from the brain. This framework employs spatial filters to enhance spatial differences between tasks while capturing both local and global energy features. We propose a novel decoding framework that combines intra-frequency and cross-frequency brain networks with GSP to extract complex energy fluctuations from the brain. This framework employs spatial filters to enhance spatial differences between tasks while capturing both local and global energy features. We show that movement information is present during preparation, and that merging the preparation features with the MI features after separate processing leads to better classification for a range of MI window lengths. We show that movement information is present during preparation, and that merging the preparation features with the MI features after separate processing leads to better classification for a range of MI window lengths. Methods Subjects A total of 14 healthy, right-handed volunteers (mean age: 24.9 ± 1.2 years; 10 men and 4 women) with normal or corrected-to-normal vision and no history of brain or upper limb disease participated in this study. All subjects, who had no prior experience with MI experiments, were postgraduate students at Hebei University of Technology. They fully understood the experimental procedures and actively cooperated throughout the study. After being informed of the study's purpose, each participant voluntarily signed a written informed consent form. This study was approved by the Biomedical Ethics Review Committee of Hebei University of Technology and was conducted in accordance with the Declaration of Helsinki (Approval No. HEBUThMEC2021007). Experimental paradigm The experimental procedure consisted of two sub-experiments: action observation-based motor imagery (AO-MI) and PMS-MI. To avoid fatigue, each subject completed the two sub-experiments over two separate days. In each sub-experiment, subjects were seated in a chair positioned 50 cm from a computer screen, with the chair's height and distance adjustable for individual comfort. On the first day, subjects performed the AO-MI experiment in two sessions. They placed their upper limbs on a table and remained relaxed while a monitor in front of them displayed the corresponding experimental instructions and a video of elbow flexion. To ensure adequate rest, a 2 min break was provided between the sessions. The specific sequence of a trial is shown in Fig. 1A. All times are reported relative to MI onset (t = 0). During the initial three seconds of each trial, the screen displayed a blank screen to maintain the subject's idle state. From – 1 s to 0 s, a white fixation cross appeared at the center of the screen, accompanied by a beep, signaling that the subject was about to begin the task. When the cross disappeared, a 4 s video of elbow flexion played, with an arrow in the middle indicating which side of the body the subject should imagine moving. While observing the arrow direction and elbow movement, subjects were instructed to simultaneously imagine the movement of the corresponding limb. At the end of the video, the word “rest” appeared on the screen, signaling a 5 s resting phase. This sequence comprised one complete trial. Each session included a total of six runs, with each run consisting of 16 trials. A 1-minute interval was provided between runs. In each set of 16 trials, videos of left and right elbow movements were shown in random order, with eight trials for each side. A total of 192 trials were recorded after completing the two AO-MI sessions. On the second day, subjects completed the second experiment. For PMS-MI, the axis is likewise aligned to t = 0 at MI onset. A 3 s idle period was provided before the start of each MI task. Unlike AO-MI, PMS-MI included an arrow on the screen initially, oriented either left or right to indicate limb orientation. The arrow remained visible for one second before transitioning into a progress bar that advanced from 0% to 100% over two seconds, during which the subject entered a prepared movement state. Upon reaching 100%, a video of the elbow flexion movement appeared on the screen, and the subject was instructed to perform the MI task in alignment with the video for four seconds. A resting phase followed, during which the subject waited for the next trial. The PMS-MI protocol was identical to AO-MI regarding the number of sessions, runs, trials, and rest intervals. The protocols differed only by the inclusion of a one-second limb orientation cue and a two-second visually guided preparation phase prior to MI (see Fig. 1B). Throughout the 2 day experiment, subjects were required to maintain their gaze at the screen center and minimize eye movements during EEG recording to reduce artifacts.Fig. 1The experimental paradigm procedure of two sub-experiments and electrodes distribution in this study. A Action observation-based MI. At the end of the cue, subjects watched the video while imagining the corresponding elbow bend movement. B Preparatory movement state-based MI. A progress bar appeared at the end of the cue to guide the subject into a prepared movement state visually. And when the progress bar disappeared, the subject began to perform the corresponding imagery. Time is expressed relative to MI onset (t = 0). The MI period lasts 4 s. C The fifty-nine electrodes used for signal acquisition and the five brain regions frontal, central, temporal, parietal and occipital are indicated by different colors The experimental paradigm procedure of two sub-experiments and electrodes distribution in this study. A Action observation-based MI. At the end of the cue, subjects watched the video while imagining the corresponding elbow bend movement. B Preparatory movement state-based MI. A progress bar appeared at the end of the cue to guide the subject into a prepared movement state visually. And when the progress bar disappeared, the subject began to perform the corresponding imagery. Time is expressed relative to MI onset (t = 0). The MI period lasts 4 s. C The fifty-nine electrodes used for signal acquisition and the five brain regions frontal, central, temporal, parietal and occipital are indicated by different colors Data recording and preprocessing EEG signals were recorded using a 64-channel portable wireless EEG amplifier (NeuSen.W64, Neuracle, China) with electrodes positioned according to the international 10/20 EEG system. As shown in Fig. 1C, fifty-nine electrodes were selected for EEG signal acquisition, covering the frontal (Fpz, Fp1, Fp2, AF3, AF4, AF7, AF8, Fz, F1, F2, F3, F4, F5, F6, F7, F8), central (FCz, FC1, FC2, FC3, FC4, FC5, FC6, Cz, C1, C2, C3, C4, C5, C6, CP1, CP2, CP3, CP4, CP5, CP6), temporal (FT7, FT8, T7, T8, TP7, TP8), parietal (Pz, P3, P4, P5, P6, P7, P8), and occipital (POz, PO3, PO4, PO5, PO6, PO7, PO8, Oz, O1, O2) regions. Five auxiliary physiological channels (HEOL, HEOR, VEOL, VEOU, ECG) record EOG/ECG rather than scalp EEG and were therefore excluded from EEG analyses. The sampling rate was set to 1000 Hz, and electrode impedance was maintained below 5 KΩ. The raw EEG data were down-sampled from 1000 Hz to 200 Hz and filtered using a finite impulse response (FIR) bandpass filter with a cut-off range of 0.1–100 Hz and a 50 Hz notch filter. Eye and muscle artifacts, along with global background noise, were removed using the independent component analysis (ICA) algorithm and common average referencing (CAR), respectively. For all subsequent analyses, the time axis was aligned to MI onset (t = 0), defined as the onset of the elbow-flexion video. Preprocessed data were segmented from −3 s to 4 s relative to MI onset. Baseline intervals were [−1, 0] s for AO-MI and [−3, −2] s for PMS-MI. All data processing was conducted in MATLAB 2020b using the open-source toolbox EEGLAB version 14.1.1 [37]. Movement-related cortical potentials (MRCPs) MRCPs are motor-related signal potentials in the low-frequency temporal band observed during ME or MI [38]. In this study, we first extracted the low-frequency temporal component from each trial using a fourth-order Butterworth filter with a frequency range of 0.1–4 Hz. Then, we averaged all trials across subjects for each task type. Event-related spectral perturbation (ERSP) To compare spectral power variations in the induced EEG concerning the stimulus within the temporal and frequency domains, the event-related spectral perturbation (ERSP) was used to map ERD/ERS in the time-frequency domain [39]. The definition of ERSP is as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ERSP\left( {f,t} \right) = \frac{1}{K}\mathop \sum \limits_{{k = 1}}^{K} \left| {F_{k} \left( {f,t} \right)} \right|^{2} $$\end{document}where K is the number of trials of each task type, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{k}(f,t)$$\end{document} denotes the spectral power estimation of the k th trial at frequency f and time t. The ERSP was baseline-normalized to the period of − 1 s to 0 s for AO-MI and − 3 s to − 2 s for PMS-MI. A Short-Time Fourier Transform (STFT) is applied for time–frequency analysis of EEG data, utilizing a Hanning-tapered window within EEGLAB [37]. MLED method The proposed method in this study, referred to as the MLED, consists of four stages. The specific process is illustrated in Fig. 2 and includes the construction of the multi layer network, the design of GFT-basis and signal projection, feature extraction using CSP and classification with a support vector machine (SVM).Fig. 2Schematic diagram of the MLED method principle. Firstly, the preprocessed EEG data were divided into five frequency bands, δ, θ, α, β and γ. The amplitude envelope correlation method was used to construct five intra-frequency coupling sub-networks, while the amplitude-amplitude coupling method was used to construct ten cross-frequency coupling sub-networks. The elements of the multi layer network matrices consist of diagonal intra-frequency network matrices and off-diagonal cross-frequency network matrices. Secondly, the filtered EEG data for different task types were vertically spliced and cross-projected with different basis functions computed by GFT to obtain GFT domain signals, respectively. Then, the CSP algorithm amplifies the difference of multichannel GFT domain signals by constructing spatial filters to obtain the feature vectors. Finally, the feature vectors of different task types are fed into the SVM classifier to obtain classification results Schematic diagram of the MLED method principle. Firstly, the preprocessed EEG data were divided into five frequency bands, δ, θ, α, β and γ. The amplitude envelope correlation method was used to construct five intra-frequency coupling sub-networks, while the amplitude-amplitude coupling method was used to construct ten cross-frequency coupling sub-networks. The elements of the multi layer network matrices consist of diagonal intra-frequency network matrices and off-diagonal cross-frequency network matrices. Secondly, the filtered EEG data for different task types were vertically spliced and cross-projected with different basis functions computed by GFT to obtain GFT domain signals, respectively. Then, the CSP algorithm amplifies the difference of multichannel GFT domain signals by constructing spatial filters to obtain the feature vectors. Finally, the feature vectors of different task types are fed into the SVM classifier to obtain classification results Multi layer network construction To constructing the multi layer brain network, the preprocessed multichannel EEG signals X were segmented into different frequency bands. In this study, the number of frequency bands \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{band}$$\end{document} was set to 5. EEG signals in the δ, θ, α, β, and γ frequency bands (δ:0.1–4 Hz, θ:4–8 Hz, α:8–12 Hz, β:12–30 Hz, and γ:30–49 Hz) were obtained using a fourth-order Butterworth filter. The intra-frequency coupling sub-networks were constructed based on AEC, while the cross-frequency coupling sub-networks were constructed using AAC. Finally, these sub-networks were constructed into a multi layer brain network matrix in the corresponding order. AAC sub-network For a given pair of EEG signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{m}(t)\in {R}^{{N}_{c}\times {N}_{t}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{n}(t)\in {R}^{{N}_{c}\times {N}_{t}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{c}$$\end{document} is the number of channels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{t}$$\end{document} is the number of time samples and m, n represent channel. Band-limited signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{{f}_{1}}^{m}(t)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{{f}_{1}}^{n}(t)$$\end{document} can be obtained after filtering with different frequency bands \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{2}$$\end{document} respectively. Next, the Hilbert Transform (HT[.]) applied to each filtered signals yields the following pair of corresponding analytic signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H(t)$$\end{document} as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{f}_{1}}^{m}\left(t\right)={x}_{{f}_{1}}^{m}\left(t\right)+jHT[{x}_{{f}_{1}}^{m}(t)]={A}_{{f}_{1}}^{m}(t){e}^{j{\varphi }_{{f}_{1}}(t)}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{f}_{2}}^{n}\left(t\right)={x}_{{f}_{2}}^{n}\left(t\right)+jHT[{x}_{{f}_{2}}^{n}(t)]={A}_{{f}_{2}}^{n}(t){e}^{j{\varphi }_{{f}_{2}}(t)}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A(t)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi (t)$$\end{document} represent the instantaneous amplitude/envelope and instantaneous phase of the signal, respectively, which are defined as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A(t)=\sqrt{x^{2}(t)+HT^{2}[x(t)]}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi \left(t\right)=arctan\frac{HT[x(t)]}{x(t)}$$\end{document} Based on the envelopes of the two signals, the AAC is estimated by calculating the following Pearson correlation formula:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${AAC}_{{f}_{1},{f}_{2}}^{m,n}=corr\left({x}_{{f}_{1}}^{m}\left(t\right),{x}_{{f}_{1}}^{n}\left(t\right)\right)=\left|\rho ({A}_{{f}_{1}}^{m}\left(t\right),{A}_{{f}_{2}}^{n}(t))\right|$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$corr(.)$$\end{document} is the function of MATLAB to calculate the Pearson correlation. Since the absolute value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} is estimated in the above equation, the AAC values are between 0 and 1, where 1 indicates a strong coupling relationship and 0 indicates no coupling relationship. Since fifty-nine electrodes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{c}$$\end{document}) were used for signal acquisition and the number of filter bands was five, ten \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{c}\times {N}_{c}$$\end{document} AAC matrices could be constructed for each trial for all subjects. AEC sub-network Unlike AAC's cross-frequency coupling for sub-networks construction, AEC is a intra-frequency coupling approach. For a given signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{m}(t)\in {R}^{{N}_{c}\times {N}_{t}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{n}(t)\in {R}^{{N}_{c}\times {N}_{t}}$$\end{document}, the two signals are filtered over the same range to obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{f}^{m}(t)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{f}^{n}(t)$$\end{document} filtered at a specific frequency band f. The corresponding analytic signal obtained after Hilbert Transform is shown below:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{f}^{m}\left(t\right)={x}_{f}^{m}\left(t\right)+jHT[{x}_{f}^{m}(t)]={A}_{f}^{m}(t){e}^{j{\varphi }_{f}(t)}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{f}^{n}\left(t\right)={x}_{f}^{mn}\left(t\right)+jHT[{x}_{f}^{n}(t)]={A}_{f}^{n}(t){e}^{j{\varphi }_{f}(t)}$$\end{document}then, the value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} of the two instantaneous amplitude under the same frequency band can be estimated by the following formula:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${AEC}_{f}^{m,n}=corr\left({x}_{f}^{m}\left(t\right),{x}_{f}^{n}\left(t\right)\right)=\left|\rho ({A}_{f}^{m}\left(t\right),{A}_{f}^{n}(t))\right|$$\end{document}like AAC, AEC has a value between 0 and 1, where 1 indicates a strong coupling relationship and 0 indicates no coupling relationship. Contrary to the number of AAC sub-networks, in the case where the number of filter bands is five, a total of five \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{c}\times {N}_{c}$$\end{document} AEC sub-network matrices under the same frequency band can be obtained. Sub-networks construction A total of 15 sub-networks were obtained by constructing the sub-networks in the intra-frequency band and cross-frequency band. The AAC sub-network constructed across the frequency bands can be represented as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M = [m_{\delta - \theta } ,m_{\delta - \alpha } ,m_{\delta - \beta } ,m_{\delta - \gamma } ,m_{\theta - \alpha } ,m_{\theta - \beta } ,m_{\theta - \gamma } ,m_{\alpha - \beta } ,m_{\alpha - \gamma } ,m_{\beta - \gamma } ]$$\end{document} Specifically, mδ-α represents the AAC sub-network between δ and θ frequency bands, mδ-α represents the AAC sub-network between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha $$\end{document} frequency bands, and so on. Typically, cross-frequency coupling connectivity is weak. To improve visualization, we have normalized the value in the matrix. For AEC sub-network constructed within the same frequency bands, it can be represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N=[{n}_{\delta },{n}_{\theta },{n}_{\alpha },{n}_{\beta },{n}_{\gamma }]$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{\delta }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{\theta }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{\alpha }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{\beta }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{\gamma }$$\end{document} denotes the AEC sub-network in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta $$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha $$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} frequency band. In each sub-network matrix, the diagonal elements are set to 0. Finally, the connectivity matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z\in {R}^{{N}_{c}{N}_{band}\times {N}_{c}{N}_{band}}$$\end{document} of the multi layer network consists of the AAC and AEC matrices as shown in Eq. (10). The diagonal elements correspond to the AEC connectivity matrices and the off-diagonal elements represent the AAC connectivity matrices.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ Z = \left[ {\begin{array}{*{20}c} {n_{\delta } } & {m_{\delta - \theta } } & {m_{\delta - \alpha } } & {m_{\delta - \beta } } & {m_{\delta - \gamma } } \\ {m_{\delta - \theta } } & {n_{\theta } } & {m_{\theta - \alpha } } & {m_{\theta - \beta } } & {m_{\theta - \gamma } } \\ {m_{\delta - \alpha } } & {m_{\theta - \alpha } } & {n_{\alpha } } & {m_{\alpha - \beta } } & {m_{\alpha - \gamma } } \\ {m_{\alpha - \beta } } & {m_{\theta - \beta } } & {m_{\alpha - \beta } } & {n_{\beta } } & {m_{\beta - \gamma } } \\ {m_{\alpha - \gamma } } & {m_{\theta - \gamma } } & {m_{\alpha - \gamma } } & {m_{\beta - \gamma } } & {n_{\gamma } } \\ \end{array} } \right] $$\end{document} Design of GFT-basis and signal projection GSP is an emerging field that builds upon graph theory, extending the classical Fourier transform into the graph domain through the GFT. By integrating principles from spectral graph theory, GSP enables the analysis of signals based on the structural relationships within graphs, leading to enhanced signal interpretation and robustness in applications. In standard terminology, a weighted graph is represented as the following triplet:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G=\{V,E,W\}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V=\left\{1,2,\dots ,\left|V\right|\right\}$$\end{document} denotes the set of nodes, E denotes the set of edges that define connectivity between nodes, and W is a weighted adjacency matrix containing connection information as well as the weight value between two nodes. The graph Laplacian, commonly used as the graph shift operator, transforms the adjacency matrix into a sparse matrix that considers the graph’s topology. The Laplacian matrix L is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L=D-W$$\end{document}, where D is the degree matrix which is a diagonal matrix. Through eigenanalysis of L, a full set of basis functions can be derived for analyzing graph signals defined on G. Eigenvectors associated with eigenvalues \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda }_{i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i=\mathrm{0,1},\dots ,\left|V\right|-1$$\end{document}, are sorted in ascending order and form a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left|V\right|\times \left|V\right|$$\end{document} matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U={[u}_{0},{u}_{1},\dots ,{u}_{\left|V\right|-1}]$$\end{document}. Projecting the original signal onto the eigenvectors of U is analogous to applying the Fourier transform in image processing, with the eigenvectors acting as orthogonal basis functions. For multichannel EEG signals X, applying the GFT involves multiplying two matrices, resulting in the signal projection:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{X}{=U}^{T}X$$\end{document} The matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{X}\in {R}^{\left|V\right|\times T}$$\end{document} contains the time-indexed projections of the multichannel signal, reflecting the time-varying contribution of the corresponding eigenmodes (T denotes the number of time samples). Given that the ERD phenomenon involves changes in power, this transformation provides richer insights into the power distribution across electrodes. It captures variations across different brain regions and frequency bands. Moreover, the time-integrated energy or power for each eigenmode or nodes can be estimated by the following formula:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Power\left(v\right)=\frac{1}{T}\sum_{t=1}^{T}{\left|\widetilde{X}(v,t)\right|}^{2}, v=\mathrm{1,2},\dots ,\left|V\right|$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Power$$\end{document} is a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left|V\right|$$\end{document}-dimensional vector and represents the distribution of signal power on different nodes. This allows the energy variations at the same nodes in different frequency bands to be more clearly displayed in the data visualization process. In our study, to obtain GFT basis functions for different task, the weighted adjacency matrix W in the Eq. (11) is represented by the multi layer network matrix Z. Therefore, the graph Laplacian matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L=D-Z$$\end{document}) can be calculated. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left|V\right|\times \left|V\right|$$\end{document} matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U={[u}_{0},{u}_{1},\dots ,{u}_{\left|V\right|-1}]$$\end{document} is the basis function of the multi layer network, where V is equal to the dimension of the multi layer network matrix (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{c}\times {N}_{band}$$\end{document}). To obtain robust GTF basis functions across subjects and tasks, we calculated the basis functions matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{U }$$\end{document} by averaging all trials and performing GFT calculations. Since the matrix Z is composed of sub-matrices of different frequency bands, in order to ensure the consistency of the data dimensions, it is necessary to adjust the EEG data, i.e. the filtered signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\delta }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\theta }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\alpha }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\beta }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\gamma }$$\end{document} are concatenated along the frequency bands to form a new matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{\delta \sim \gamma }\in {R}^{{N}_{c}{N}_{band}\times {N}_{t}}$$\end{document}. After the adjustment, the dimension of the data changes from the number of channels under a single band to the product of multiple bands and the number of channels. Finally, the time-indexed projections of the multichannel GFT domain signal can be calculated by the following formula:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{GFT}={\overline{U} }^{T}{X}^{\delta \sim \gamma }$$\end{document} For different task type, we can obtain different basis functions matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{U }$$\end{document}, such as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{2}$$\end{document}. Feature extraction The CSP algorithm is a spatial filtering method extensively applied in MI-BCI feature extraction. Its primary function is to identify spatial filters that maximize variance for one task category while minimizing it for another, effectively acting as a difference amplifier as well as a spatial filter. In this study, the CSP’s objective is to construct optimal projections that enhance the variance distinction of multichannel signals in the GFT domain for two task types. For two different task types of signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{1}^{\delta \sim \gamma }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{2}^{\delta \sim \gamma }$$\end{document}, two kinds of basis functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{2}$$\end{document} can be obtained after GFT calculation. Since each type of signal needs to face different basis functions, our solution is to project each type of signal with different basis functions separately to obtain four GFT domain signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{1}^{T}{X}_{1}^{\delta \sim \gamma }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{2}^{T}{X}_{1}^{\delta \sim \gamma }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{1}^{T}{X}_{2}^{\delta \sim \gamma }$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{U} }_{2}^{T}{X}_{2}^{\delta \sim \gamma }$$\end{document}. Then the four projected signals are spliced along the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{t}$$\end{document} direction to obtain new two task types GFT domain signals, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\dot{X}}_{1}=[{\overline{U} }_{1}^{T}{X}_{1}^{\delta \sim \gamma },{\overline{U} }_{2}^{T}{X}_{1}^{\delta \sim \gamma }]\in {R}^{{N}_{c}{N}_{band}\times 2{N}_{t}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\dot{X}}_{2}=[{\overline{U} }_{1}^{T}{X}_{2}^{\delta \sim \gamma },{\overline{U} }_{2}^{T}{X}_{2}^{\delta \sim \gamma }]\in {R}^{{N}_{c}{N}_{band}\times 2{N}_{t}}$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\dot{X}$$\end{document} represents the GFT domain signal after splicing for a specific task type. The covariance matrix is then represented as follows:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C=\frac{\dot{X}{\dot{X}}^{T}}{trace(\dot{X}{\dot{X}}^{T})}$$\end{document} Assume that the training trials contain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{i}$$\end{document} task types, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i\in \{\mathrm{1,2},\dots ,K\}$$\end{document} denotes the subject’s task labels. Each task contains \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{i}$$\end{document} GFT domain signals, and the average covariance of each task \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{i}$$\end{document} is calculated as follows:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{i}=\frac{1}{{N}_{i}}\sum_{C\in {K}_{i}}C$$\end{document} The spatial filters between task types \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{j}$$\end{document} can be calculated as follows:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J(\omega )=\frac{{\omega }^{T}{R}_{i}\omega }{{\omega }^{T}{R}_{j}\omega }$$\end{document} Equation (17) can be solved by generalized eigenvalue decomposition. These eigenvectors with correspond to different generalized eigenvalues. The optimal spatial filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{CSP}=[{\omega }_{1},{\omega }_{2},\dots ,{\omega }_{2E}]$$\end{document} is obtained by sorting the eigenvectors and selecting those corresponding to the E maximum and minimum generalized eigenvalues. Using these filters, the CSP feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f={[{f}_{1},{f}_{2},\dots ,{f}_{2E}]}^{T}$$\end{document} is calculated as follows:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{e}=\mathrm{log}\left(var\left({\omega }_{e}^{T}\dot{X}\right)\right),e=\mathrm{1,2},\dots ,2E$$\end{document} Classification Following the extraction of CSP feature vectors, SVM classifier was employed for task classification. SVM, a widely used supervised machine learning method, is well-suited for EEG feature classification. In this study, CSP feature vectors were input into an SVM classifier with a radial basis function (RBF) kernel to distinguish between different tasks. The RBF kernel SVM classifier was trained using MATLAB 2020b’s fitcsvm function, with default kernel parameters. Classification accuracy for each subject was evaluated through a 10×10-fold cross-validation approach, in which 90% of the data served as the training set, and 10% as the test set, with the process repeated ten times to ensure robust performance assessment. Comparison algorithms To enable a more detailed comparison, this study covers the following commonly used MI EEG signal classification algorithms: (1) FBCSP. Built on CSP, FBCSP first applies a filter bank to split EEG into multiple sub-bands, extracts CSP features per band, performs feature selection, and then classifies the selected features. To keep our pipeline consistent, we used an SVM classifier on the FBCSP features [24]. (2) MDM (Minimum Distance to Riemannian Mean). Each trial is represented by its spatial covariance (an SPD matrix). For each class, a Riemannian (affine-invariant) mean is computed on the SPD manifold; a test trial is assigned to the class whose mean has the smallest Riemannian distance to the trial covariance. The approach is simple and robust, particularly under limited training data [40]. (3) sCNN (ShallowConvNet). A shallow EEG-specific ConvNet that stacks temporal convolution and spatial (channel) convolution, followed by pooling and dropout; it targets sensorimotor rhythms with few layers and parameters [41]. (4) dCNN (DeepConvNet). A deeper ConvNet with multiple temporal–spatial convolution and pooling stages for hierarchical feature extraction from raw EEG, typically paired with dropout for regularization [41]. (5) EEGNet. A compact ConvNet tailored to EEG that uses depthwise and separable convolutions to decouple spatial and temporal filtering, achieving strong performance with very few parameters [42]. (6) FBCNet. A multi-view deep model that first builds a filter-bank (multi-band) representation, then learns spectro-spatial features via convolution and a spatial transform, and finally classifies with fully connected layers [43]. All deep learning baselines (sCNN, dCNN, EEGNet and FBCNet) were trained in PyTorch using the Adam optimizer with an initial learning rate of 0.01 and a cosine-annealing learning-rate scheduler. Training ran for up to 200 epochs with early stopping when the validation loss showed no improvement for 50 consecutive epochs; the best validation checkpoint was retained for testing. The batch size was 32 for all baselines. Unless otherwise stated, no weight decay was used and the cross-entropy loss optimized all models. Architecture-specific regularization followed their canonical designs: EEGNet used dropout = 0.25, while sCNN and dCNN used dropout = 0.5; FBCNet employed batch normalization and max-norm constraints (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Vert w\Vert }^{2}$$\end{document}< 2 in the spatial convolution block and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Vert w\Vert }^{2}$$\end{document} < 0.5 in the final fully connected layer) without dropout. Results AO-MI representation We analyzed low-frequency MRCPs and high-frequency ERD features of AO-MI to compare the classical and proposed paradigms. Fig. 3A shows average MRCPs waveforms for left and right movements on electrodes FC1 and FC2. At time zero, both movements showed increased amplitudes on FC1 and FC2. Around 700 ms, the amplitude shifted rapidly from positive to negative. On the FC1 electrode, left movement amplitude was lower than right movement amplitude from 0 to 700 ms. However, after the rapid decrease at 700 ms, the amplitude of the left movement became higher than that of the right movement; this phenomenon was reversed on the FC2 electrode. To observe changes occurring in the brain throughout the process more intuitively, topographic maps were plotted from 500 ms to 1000 ms for both task types. The results indicate that in the central and prefrontal regions, the topographic maps under different tasks exhibited a symmetrical pattern characterized by spatial differences (Fig. 3B). Fig. 3C shows average time-frequency maps of high-frequency ERD features on C3 and C4. EEG power showed clear alpha and beta ERD features within 4 seconds of MI onset. Fig. 3D further illustrates that ERD features, induced by different motor tasks in the alpha, beta, and combined alpha+beta bands, show a clear contralateral dominance. For instance, in the alpha band, brain topography for left movements demonstrated bilateral ERD, with more pronounced effects in the contralateral regions. Additionally, differences in ERD between left and right tasks were computed to highlight the spatial distinction of ERD features across tasks. The topography of these differences revealed a stronger contralateral distribution, with higher energy on the left and lower energy on the right. To quantitatively support these spatial patterns, Fig. 3E summarizes the mean ERD at C3/C4 across frequency bands. For left-hand imagery, C3/C4 means were −0.91/−1.29 (alpha), −0.42/−0.70 (beta), and −0.55/−0.85 (alpha+beta). For right-hand imagery, the corresponding C3/C4 means were −1.27/−0.60 (alpha), −0.58/−0.33 (beta), and −0.76/−0.40 (alpha+beta). Consistent with the topographies, these values show a clear contralateral dominance. Together with the difference maps, these numeric contrasts confirm robust contralateral desynchronization—most pronounced in the beta and combined bands—while alpha ERD remains bilateral but contralaterally stronger.Fig. 3The low-frequency MRCPs features and high-frequency ERD patterns of AO-MI paradigm. A The average waveforms of MRCPs across all subjects of left and right movements on electrodes FC1 and FC2. B The spatial distribution of average MRCPs ranges from 500 to 1000 ms corresponding to left and right movements tasks. C The averaged time–frequency maps on electrodes C3 and C4 for left and right movement tasks. D Brain power topography in alpha, beta, and alpha + beta bands for left movement, right movement, and differences between these two tasks. (E) Averaged ERD value of C3 and C4 in three frequency bands. The error bars represent the standard error of the mean The low-frequency MRCPs features and high-frequency ERD patterns of AO-MI paradigm. A The average waveforms of MRCPs across all subjects of left and right movements on electrodes FC1 and FC2. B The spatial distribution of average MRCPs ranges from 500 to 1000 ms corresponding to left and right movements tasks. C The averaged time–frequency maps on electrodes C3 and C4 for left and right movement tasks. D Brain power topography in alpha, beta, and alpha + beta bands for left movement, right movement, and differences between these two tasks. (E) Averaged ERD value of C3 and C4 in three frequency bands. The error bars represent the standard error of the mean PMS-MI representation Under PMS-MI, group-average MRCPs at FC1/FC2 were small in amplitude and bilaterally comparable, with values mostly within ±1 μV (about −1 to +1 μV) and no consistent canonical negative deflection or lateralization. However, for subject S12, spatial differences in MRCPs patterns were evident, with contralateral electrodes displaying lower amplitudes and greater rates of decline during left or right movements (Fig. 4A). For the averaged time-frequency maps on electrodes C3 and C4, distinct ERD phenomena were observed. As shown in Fig. 4B, during the preparatory movement period (− 2 s to 0 s), energy decreases occurred near the 5–10 Hz bands on electrodes C3 and C4 across different task types. During the four seconds of MI, ERD was also evident in the alpha and beta bands across tasks at both electrodes, with a more dispersed energy pattern compared to AO-MI. Additionally, unlike AO-MI, a reduction in energy was also observed in lower frequency bands (delta and theta). During the MI window, ERD became stronger and the spatial patterns were clearer. Consistent with the topographies in Fig. 4C, the averaged ERD at C3/C4 (Fig. 4E) was as follows. For left-hand imagery, C3/C4 were −1.02/−1.00 in the alpha band, −0.43/−0.67 in the beta band, and −0.58/−0.76 in the combined alpha+beta band. For right-hand imagery, C3/C4 were 0.87/−0.87 (alpha), −0.49/−0.69 (beta), and −0.59/−0.73 (alpha+beta). These quantitative values accompany the maps in Fig. 4C, showing broad-band desynchronization during PMS-MI with band-specific lateralization. To characterize the preparatory period, Fig. 4D depicts largely bilateral patterns, which are complemented by the C3/C4 means in Fig. 4E. For left-hand imagery, C3/C4 were −0.57/−0.36 (under_alpha), 0.10/0.20 (beta), and −0.07/0.08 (under_alpha+beta). For right-hand imagery, C3/C4 were −0.56/−0.77 (under_alpha), 0.23/0.06 (beta), and 0.07/−0.09 (under_alpha+beta). These values indicate weaker modulation during preparation with mixed ERD/ERS across bands, aligning with the near-symmetric topographies in Fig. 4D.Fig. 4The low-frequency MRCPs features and high-frequency ERD patterns of PMS-MI paradigm. A The average waveforms of MRCPs across all subjects and individual S12 of left and right movements on electrodes FC1 and FC2. B The averaged time–frequency maps on electrodes C3 and C4 for left and right movement tasks for the time range from –2 s to 4 s, including the preparation movement period and the MI period. C Brain power topography in alpha, beta, and alpha + beta bands for left movement, right movement, and differences between these two tasks for 0 s-4 s (MI period). D Brain activity in under_alpha, beta, and under_alpha + beta bands for left and right tasks during the 2 s period before movement begins. E Averaged ERD value of C3 and C4 in three frequency bands. The shaded area represents data from − 2 s to 0 s. The unshaded area represents data from 0 to 4 s. The error bars represent the standard error of the mean The low-frequency MRCPs features and high-frequency ERD patterns of PMS-MI paradigm. A The average waveforms of MRCPs across all subjects and individual S12 of left and right movements on electrodes FC1 and FC2. B The averaged time–frequency maps on electrodes C3 and C4 for left and right movement tasks for the time range from –2 s to 4 s, including the preparation movement period and the MI period. C Brain power topography in alpha, beta, and alpha + beta bands for left movement, right movement, and differences between these two tasks for 0 s-4 s (MI period). D Brain activity in under_alpha, beta, and under_alpha + beta bands for left and right tasks during the 2 s period before movement begins. E Averaged ERD value of C3 and C4 in three frequency bands. The shaded area represents data from − 2 s to 0 s. The unshaded area represents data from 0 to 4 s. The error bars represent the standard error of the mean Classification Previous analyses of the time and frequency domain representations of EEG data revealed energy variations and spatial differences in the brain during different tasks. This phenomenon forms the basis for feature extraction and classification. Table 1 presents the average classification results across paradigms and window lengths using different algorithms. The results indicate that under the traditional AO-MI paradigm, the MLED method achieved the highest classification accuracy across all window lengths. The FBCNet method demonstrated greater stability compared to other deep learning methods used in this study and generally outperformed them in classification accuracy. For the PMS-MI paradigm, the MLED method achieved the highest classification accuracy only at the 0–1 s window length. At other window lengths, its accuracy was slightly lower than deep learning methods but still outperformed other machine learning approaches. Multidimensional comparisons were conducted using statistical analysis methods: one-way ANOVA with Tukey correction for different algorithms within the same paradigm, and paired t-tests for the same algorithm across different paradigms. Results indicate that the MLED method showed no statistically significant differences across different paradigms for window lengths of 0–1 s and 0–3 s. For the 0–3 s window length, all machine learning methods exhibited no statistically significant differences across different paradigms. Given that the time-frequency analysis of the PMS-MI paradigm revealed similar energy changes during the preparation movement period, we conducted a classification study on the data from this period (− 2 to 0 s), as shown in Fig. 5A. Using the CSP and MDM method, classification accuracies of 67.14% and 69.12% were obtained, though this level was insufficient to significantly enhance BCI system performance. The FBCSP method achieved a classification accuracy of 74.53%, which has reached the basic level required for using a BCI system. The remaining deep learning methods further improved classification accuracy. Among them, the FBCNet method achieved 78.40%, which was 1.25% lower than the MLED method (one-way ANOVA with Tukey correction, F(7, 104) = 4.883, p < 0.05). Subsequently, we attempted to decode data from both the preparation movement period and the MI period continuously. The analysis of spatial brain topography during the preparation period revealed varying frequency ranges and spatial differences, indicating that effective extraction of both signal types requires distinct approaches. Therefore, to deepen understanding and enhance decoding accuracy within this paradigm, we applied the MLED method to visualize features during both the preparation and MI periods. As shown in Fig. 5B, the feature distributions for the two periods exhibit minimal overlap and are clearly separable. Finally, Fig. 5C presents the classification results achieved using different decoding processing approaches for 2 s movement preparation time and various MI window lengths. It can be observed that the segmented decoding method consistently outperformed the continuous decoding method with all algorithms. For the segmented decoding method with deep learning, we used an end-to-end two-branch deep model to mirror the classic “extract features from two segments, then fuse and classify” pipeline. Each branch encoded one EEG segment with the same backbone (sCNN, dCNN, EEGNet, or FBCNet). The two representations were concatenated (late feature fusion) and fed to a lightweight classifier. For the basic CSP algorithm, the segmented approach yielded the highest accuracy improvement compared to continuous decoding methods. For machine learning methods such as MDM and FBCSP, the segmented approach also yielded significantly enhanced classification accuracy. Deep learning methods demonstrated gradually increasing and stable classification accuracy in continuous decoding as the decoding window length increased. The MLED method achieves a classification accuracy of 79.69% at a window length of -2-1 s, outperforming all other algorithms. However, as the analysis window length increases, its classification accuracy gradually declined and fell below that of deep learning methods, maintaining an accuracy of 76.91% at -2-4 s. For segmented decoding methods, MLED demonstrates a competitive advantage. Specifically, with 1 s and 4 s MI window lengths, the segmented decoding approach achieved decoding accuracies of 86.56% and 86.13%, respectively, representing improvements of 6.87% (p < 0.05, paired t-test) and 9.22% (p < 0.05, paired t-test) compared to the continuous decoding method. For 2 s and 3 s window lengths, the decoding accuracy of the segmented approach was slightly lower, achieving 84.01% and 84.27%, respectively. However, these values still showed improvements of 7.06% (p < 0.05, paired t-test) and 9.59% (p < 0.05, paired t-test) over the continuous decoding method.Table 1Mean ± SD (%) classification accuracy of eight algorithms—CSP, MDM, FBCSP, sCNN, dCNN, EEGNet, FBCNet, and MLED—across two paradigms (AO-MI, PMS-MI) and four decoding windows (0–1 s, 0–2 s, 0–3 s, 0–4 s). Within each paradigm, a one-way ANOVA with Tukey correction tested the main effect of algorithm on accuracyWindowParadigmAlgorithmsOne-way ANOVACSPMDMFBCSPsCNNdCNNEEGNetFBCNetMLED0–1 sAO-MI68.87 ± 7.4769.12 ± 8.7563.63 ± 2.5067.93 ± 4.3967.55 ± 5.0865.81 ± 6.8666.55 ± 6.7482.28 ± 5.73F(7,104) = 11.62 < 0.05PMS-MI76.76 ± 10.4875.06 ± 10.2975.72 ± 8.3377.65 ± 6.9577.74 ± 5.6575.81 ± 8.4978.33 ± 5.2580.77 ± 7.77F(7,104) = 1.113 nsPaired t-test < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05ns–0–2 sAO-MI65.63 ± 5.5767.82 ± 8.1368.03 ± 5.7072.22 ± 6.6371.20 ± 6.9969.67 ± 6.7774.47 ± 6.4383.11 ± 6.13F(7,104) = 9.568 < 0.05PMS-MI71.54 ± 11.8275.21 ± 8.8574.06 ± 9.2481.80 ± 8.2982.80 ± 6.7680.39 ± 9.0281.23 ± 7.6275.80 ± 8.34F(7,104) = 3.145 < 0.05Paired t-testns < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05–0–3 sAO-MI67.90 ± 11.8468.75 ± 9.5072.25 ± 8.8173.63 ± 9.5675.51 ± 9.4672.86 ± 6.8175.52 ± 7.7381.47 ± 7.00F(7,104) = 3.170 < 0.05PMS-MI72.11 ± 13.5074.82 ± 11.5775.79 ± 9.9481.18 ± 7.9680.67 ± 8.0580.36 ± 8.3782.24 ± 7.4575.16 ± 9.40F(7,104) = 2.073 nsPaired t-testnsnsns < 0.05 < 0.05 < 0.05 < 0.05ns–0–4 sAO-MI67.17 ± 9.1167.73 ± 10.5173.12 ± 8.3674.31 ± 9.6976.03 ± 9.1774.89 ± 9.0777.69 ± 9.0881.84 ± 4.92F(7,104 = 4.230 < 0.05PMS-MI69.35 ± 10.5174.50 ± 11.3777.72 ± 9.1382.48 ± 7.8282.29 ± 6.8381.34 ± 9.9282.17 ± 7.5379.88 ± 7.78F(7,104) = 3.798 < 0.05Paired t-testns < 0.05 < 0.05 < 0.05 < 0.05<0.05nsns–The Paired t-test row reports contrasts PMS-MI vs AO-MI for each algorithm. ns = not significantFig. 5Classification results and feature distribution. A Classification results during preparation movement period (− 2 s–0 s) using different methods. The black dashed line represents the mean value. B The feature distribution of preparation and MI period. C The classification performance using different method with continuous and segmented decoding processing method at 2 s preparation time combined with four MI window lengths Table 1Mean ± SD (%) classification accuracy of eight algorithms—CSP, MDM, FBCSP, sCNN, dCNN, EEGNet, FBCNet, and MLED—across two paradigms (AO-MI, PMS-MI) and four decoding windows (0–1 s, 0–2 s, 0–3 s, 0–4 s). Within each paradigm, a one-way ANOVA with Tukey correction tested the main effect of algorithm on accuracyWindowParadigmAlgorithmsOne-way ANOVACSPMDMFBCSPsCNNdCNNEEGNetFBCNetMLED0–1 sAO-MI68.87 ± 7.4769.12 ± 8.7563.63 ± 2.5067.93 ± 4.3967.55 ± 5.0865.81 ± 6.8666.55 ± 6.7482.28 ± 5.73F(7,104) = 11.62 < 0.05PMS-MI76.76 ± 10.4875.06 ± 10.2975.72 ± 8.3377.65 ± 6.9577.74 ± 5.6575.81 ± 8.4978.33 ± 5.2580.77 ± 7.77F(7,104) = 1.113 nsPaired t-test < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05ns–0–2 sAO-MI65.63 ± 5.5767.82 ± 8.1368.03 ± 5.7072.22 ± 6.6371.20 ± 6.9969.67 ± 6.7774.47 ± 6.4383.11 ± 6.13F(7,104) = 9.568 < 0.05PMS-MI71.54 ± 11.8275.21 ± 8.8574.06 ± 9.2481.80 ± 8.2982.80 ± 6.7680.39 ± 9.0281.23 ± 7.6275.80 ± 8.34F(7,104) = 3.145 < 0.05Paired t-testns < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05 < 0.05–0–3 sAO-MI67.90 ± 11.8468.75 ± 9.5072.25 ± 8.8173.63 ± 9.5675.51 ± 9.4672.86 ± 6.8175.52 ± 7.7381.47 ± 7.00F(7,104) = 3.170 < 0.05PMS-MI72.11 ± 13.5074.82 ± 11.5775.79 ± 9.9481.18 ± 7.9680.67 ± 8.0580.36 ± 8.3782.24 ± 7.4575.16 ± 9.40F(7,104) = 2.073 nsPaired t-testnsnsns < 0.05 < 0.05 < 0.05 < 0.05ns–0–4 sAO-MI67.17 ± 9.1167.73 ± 10.5173.12 ± 8.3674.31 ± 9.6976.03 ± 9.1774.89 ± 9.0777.69 ± 9.0881.84 ± 4.92F(7,104 = 4.230 < 0.05PMS-MI69.35 ± 10.5174.50 ± 11.3777.72 ± 9.1382.48 ± 7.8282.29 ± 6.8381.34 ± 9.9282.17 ± 7.5379.88 ± 7.78F(7,104) = 3.798 < 0.05Paired t-testns < 0.05 < 0.05 < 0.05 < 0.05<0.05nsns–The Paired t-test row reports contrasts PMS-MI vs AO-MI for each algorithm. ns = not significant Mean ± SD (%) classification accuracy of eight algorithms—CSP, MDM, FBCSP, sCNN, dCNN, EEGNet, FBCNet, and MLED—across two paradigms (AO-MI, PMS-MI) and four decoding windows (0–1 s, 0–2 s, 0–3 s, 0–4 s). Within each paradigm, a one-way ANOVA with Tukey correction tested the main effect of algorithm on accuracy F(7,104) = 9.568 < 0.05 F(7,104) = 3.145 < 0.05 F(7,104) = 3.170 < 0.05 The Paired t-test row reports contrasts PMS-MI vs AO-MI for each algorithm. ns = not significant Classification results and feature distribution. A Classification results during preparation movement period (− 2 s–0 s) using different methods. The black dashed line represents the mean value. B The feature distribution of preparation and MI period. C The classification performance using different method with continuous and segmented decoding processing method at 2 s preparation time combined with four MI window lengths Network pattern Time-frequency features alone cannot effectively and comprehensively capture the strength of interactions and connections between different brain regions. Therefore, we visualized brain network patterns using sub-network matrices to provide a more comprehensive view of brain network features. Statistical analyses (paired t-test) were performed on intra-frequency and cross-frequency sub-network matrices across different tasks, with statistically significant matrices set to 1 and non-significant matrices set to 0. These statistically significant matrices were then visualized as brain networks using the BrainNet Viewer [44]. The single-layer network results for intra-frequency and cross-frequency bands are provided in Supplementary Material (Fig. S1, Fig. S2, and Fig. S3). In this study, we visualized brain networks as sub-networks for intra-frequency, cross-frequency, and a combined network representing the summation of both. The colorbar value indicates the total connection strength between electrode pairs across all sub-networks. As shown in Fig. 6A, under the AO-MI paradigm, connections are observed across all brain regions in both intra-frequency and cross-frequency bands. After threshold filtering, prominent connections emerge between the frontal and central regions in the intra-frequency band, while connections in the cross-frequency band cover the entire brain. Thresholding the intra- and cross-frequency sub-networks further highlights a strong connectivity between the frontal and central regions. Fig. 6B illustrates brain network characteristics during the preparation period of the PMS-MI paradigm. During this period, the intra-frequency band shows connectivity primarily between the prefrontal and motor regions, while cross-frequency band connectivity spans the whole brain. During the MI period, intra-frequency connections are concentrated between the prefrontal and motor regions, with some connectivity extending to the visual area. In contrast, the cross-frequency band maintains strong connectivity across the whole brain. Network thresholds for the combined intra- and cross-frequency bands reveal interactions between the frontal and motor areas, as well as connections extending into the visual regions (see Fig. 6C).Fig. 6Statistically significant patterns of brain network connectivity. A Statistically significant brain networks for intra-frequency summation, cross-frequency summation and summation of both in the AO-MI paradigm. B and C Statistically significant brain networks for intra-frequency summation, cross-frequency summation and summation of both in the PMS-MI paradigm during preparation movement period and MI period. The colorbar represents the total connection or statistically strength of the current electrode pair across all sub-networks Statistically significant patterns of brain network connectivity. A Statistically significant brain networks for intra-frequency summation, cross-frequency summation and summation of both in the AO-MI paradigm. B and C Statistically significant brain networks for intra-frequency summation, cross-frequency summation and summation of both in the PMS-MI paradigm during preparation movement period and MI period. The colorbar represents the total connection or statistically strength of the current electrode pair across all sub-networks Discussion Previous studies have demonstrated that the preparatory movement state can enhance EEG representations of premovement activity. Motor-BCI are typically based on motor imagery (MI) in addition to premovement signals. However, previous work has not fully examined the impact of the preparatory movement state on MI, which this study aims to fill this gap. Although existing literature confirms that both ME and MI can induce ERD and MRCPs features, research also indicates distinct underlying mechanisms between the two. Therefore, to achieve more accurate experimental results, we opted to induce neural oscillation changes in the brain using MI rather than ME. In this study, we designed an MI paradigm that incorporates the preparatory movement state and compared it with the classical AO-MI paradigm. We conducted neural correlation analyses of MRCPs, event-related oscillations, brain network patterns, and employing the MLED algorithm for motor decoding. MRCPs Numerous studies have shown that MRCPs is a time-domain potential feature present during both movement preparation and execution, with the contralateral hemisphere of the limb typically dominating activity during movement [45]. As observed in Fig. 3A, elbow movements exhibit distinct MRCPs features in the precentral and frontal regions, differing from lower limb movements and from the time-domain waveforms produced by finger tapping, likely due to differences in experimental paradigms [46, 47]. Analysis of the time-domain curves reveals a pattern in which the potentials rise before declining. This phenomenon may be explained by the visual stimuli presented within the paradigm: when the cue concludes and screen content changes, the moment of switching creates a visual stimulus for the subjects. The time-frequency results in Fig. 3C show an increase in energy within the low-frequency band starting from moment zero, consistent with previous findings in the literature [13]. In contrast to the AO-MI paradigm, the PMS-MI paradigm does not exhibit a clear MRCPs feature in the low-frequency time-domain average, except in Subject 12. As MI is a mental rehearsal process rather than an actual movement, it takes some time for the effects of MI to emerge. Additionally, variability in MRCP onset arises from the paradigm’s movement preparation requirements and from each subject’s intrinsic start time. As a result, the averaged waveform becomes smoothed, obscuring reliable time-domain information. The preparatory movement state may also affect the baseline of MRCPs, contributing to this variability. Whether this paradigm consistently produces MRCPs features will be further explored in the following section on brain networks. ERD/S oscillations ERD and ERS energy modulations are well-established phenomena in MI. They appear in both the alpha and beta bands under the classical AO-MI paradigm, exhibiting clear contralateral dominance in the central brain regions. Comparing left and right tasks, the energy in the left brain region exceeds that of the right, providing a theoretical basis for the classical CSP algorithm. In contrast to the contralateral ERD/ERS oscillations typically observed in the classical paradigm, the PMS-MI paradigm exhibits no clear lateralization in alpha-band topography during the 0–4 s MI period. The activation of bilateral regions appears largely symmetrical. This finding resembles the pattern seen in simultaneous bilateral limb MI, where bilateral ERDs emerge due to the heightened complexity from motor readiness [48]. However, this bilateral activation may primarily reflect intracortical feedback within MI. Such feedback modulation could synchronize ERD phenomena across both hemispheres, reflecting the brain’s mechanisms for motor preparation and execution [49]. In the beta band, brain topography reveals stronger contralateral activation during left movements, consistent with findings in the literature. However, for right movements, contralateral activation appears weaker than ipsilateral activation. Since the beta band is linked to movement execution and maintenance, and subjects were right-handed, contralateral brain regions might efficiently handle the task with less preparatory activation. This phenomenon could suggest a hemispheric competition mechanism in which the brain increases ipsilateral cortex activity to compensate for reduced contralateral activation [50]. Comparative analyses also reveal spatial energy differences across frequency bands: contrary to traditional paradigm results, energy in the left brain region is lower than in the right. Time-frequency maps during the – 2 s to 0 s motor preparation period show an ERD-like energy reduction below the alpha band. During this period, brain topography shows almost symmetrical activation of the motor areas, with maximal activity at Cz. This site, located over the sensorimotor cortex, is known to be essential for motor control and sensory integration. The strong correlation of this region with motor planning and execution [51] suggests its involvement here. Given the overlap of this low-frequency band with the MRCPs frequency range, these energy changes impact the MRCPs baseline, providing a more comprehensive explanation of the MRCPs time-domain features. The central ERS may reflect the formation of a bottom-up pathway for retrieving motor-related experiences stored in the sensorimotor cortex, aiding in interpreting perceived information and preparing for simulated movement [52, 53]. Consequently, the central ERS could indicate a heightened need for attention and vigilance during movement preparation. Difference analysis results during MI align with these observations, and this energy variance could serve as the basis for subsequent feature extraction and classification. Network and classification Time-domain and frequency-domain features typically reflect interactions between brain regions within a single frequency band, yet these features alone may not offer a fully comprehensive neural representation. Brain networks, however, can capture connectivity across diverse brain regions effectively. In this study, the single-layer network matrix represents local connectivity features, while combining matrices across different frequency bands yields a global representation. Designing effective decoding algorithms based on these features can optimize system performance, resulting in a synergy of neuroscience and computer science rather than relying solely on black-box operations. We observe strong connectivity between brain regions in the delta band within the AO-MI paradigm in SFig. 1 (Supplementary file) alongside Fig. 6A. Movement initiation and planning involve distributed engagement of sensorimotor and premotor/SMA networks, characterized by β-band desynchronization (ERD) and occasional ERS. During movement execution and maintenance, however, the beta band engages more intensively to support ongoing movement processes [54]. Previous studies have shown that the prefrontal cortex is actively involved in movement tasks. Furthermore, the sensorimotor regions not only control isolated movements but also interact with other motor-related areas during motor imagery, leading to enhanced neural coupling across these networks [55, 56]. SFig. 2 reveals close beta-band interactions between brain regions during movement preparation, complementing the ERS activity observed in motor regions described previously. Mechanistic interpretations are provided in Section "Neurobiological interpretation of intra-/inter-layer connectivity". Threshold filtering reveals prominent engagement of the frontal and occipital regions during this phase. These findings imply that visual areas contribute to higher-order cognitive processes for interpreting visual cues, thereby facilitating movement characterization and planning [57]. Following sufficient motor preparation, delta and beta bands exhibit strong inter-regional connectivity during motor imagery, mirroring the classical paradigm results (see SFig. 3). However, threshold filtering highlights interactions between frontal and visual regions, in addition to motor regions, which may imply that visual inputs initiated during preparation persist to provide support throughout the task. The algorithm builds upon the energy modulations induced by MI, focusing on brain network dynamics as its core. It analyzes network characteristics across different stages and paradigms to complement and interpret the classification outcomes. During movement preparation, the brain allocates more resources, which reduces available resources during MI. After threshold screening, the network analysis revealed fewer and weaker connections in the PMS-MI paradigm than in the AO-MI paradigm. As a result, the classification performance of MLED was inferior to that of other methods, except within the 0–1 s window. The improved classification accuracy in the PMS-MI paradigm using the traditional machine learning may be attributed to the broader energy distribution and higher degree of discretization within frequency bands. This distribution, when combined with inherent spatial differences, allows them to extract richer energy-related information (or discriminative information), thereby enhancing classification accuracy. Additionally, it was observed that the phenomenon of energy reduction occurred earlier in PMS-MI than in the AO-MI paradigm, as further corroborated by the classification results obtained from short time windows. This finding suggests that PMS-MI enhances neural representations of high-frequency ERD within single frequency bands. Such enhancement is not numerical, as seen in signal-to-noise ratio improvements, but rather temporal and spatial, involving earlier energy changes and a broader distribution range within frequency bands. However, from a global perspective, since movement preparation consumes part of the brain's resources, a reduction in connectivity strength is observed, reflecting a diminished neural representation. This indirectly highlights that the brain's resource allocation is inherently limited, unable to distribute additional resources across multiple dimensions to universally enhance cognition or memory and learning capabilities. The classification results and statistical analysis further indicate that algorithmic differences within the AO-MI paradigm introduced variability in performance. Conversely, when using different algorithms on the PMS-MI paradigm, no significant differences emerged, particularly under the 0–1 s and 0–3 s window conditions. This may suggest that performance in the early and mid-to-late stages of this paradigm is relatively stable. Moreover, under the 0–3 s window condition, multiple instances were observed where the same algorithm did not exhibit differences across different paradigms. In sum, both algorithm and paradigm matter, but their effects are condition-dependent, with several contrasts small or non-significant outside specific configurations. Comparing brain networks between the motor preparation and MI phases reveals distinct neural activity patterns, further supported by results in Fig. 5B and 5C, which highlight separable feature distributions for the two stages. By segmently extracting features from both phases and subsequently fusing them, the MLED method achieved classification accuracies of 86.56%, 84.01%, 84.27%, and 86.13% across four window lengths. Since movement planning and memorization during the preparation phase require greater resource allocation, decoding performance in the preparation phase (− 2 s to 0 s) is nearly equivalent to that in the MI phase (0 s to 4 s). During the first second of the 4 s MI period, premovement-related MRCPs features are predominantly present. This phase exhibits classification results comparable to those observed in both the preparation and MI phases. These findings indicate that the MLED method effectively extracts the brain's state-specific features across different phases. However, when features from multiple states are fused, performance tends to decline. We hypothesize that this mechanism stems from the differing statistical properties between the preparation and MI phases. Directly concatenating all data into a single long window can lead to distributional confounding. By contrast, a segmented modeling strategy with decision-layer fusion better retains complementary information from both phases, thereby mitigating the risk of mutual “dilution”. This indicates that continuous decoding cannot simultaneously capture all unique neural representations comprehensively, highlighting the necessity of tailoring feature extraction for specific stages. In terms of algorithm adaptation, this explains why MLED performs better in short-window/segmented settings, while deep networks demonstrate greater advantages in long-window/end-to-end aggregated contexts. These findings suggest that algorithms and window lengths can be tailored to specific application scenarios, thereby optimizing system robustness and flexibility. Neurobiological interpretation of intra-/inter-layer connectivity Building on the network patterns summarized in Section "Network and classification", we relate the multilayer connectivity to established circuitry for MI and preparation. Within-band (AEC): during preparation, enhanced alpha-band edges linking frontal and parietal nodes are consistent with fronto-parietal/dorsal-attention systems that impose top-down priors on sensorimotor processing [58]; during MI, beta-band edges concentrate over sensorimotor cortex together with SMA/premotor regions, in line with canonical mu/beta dynamics associated with motor set and maintenance [59, 60]. Cross-band (AAC) coupling between alpha and beta rhythms connects parietal and fronto-parietal regions with central sensorimotor nodes. This coupling likely reflects a gating-and-integration mechanism, where alpha-mediated attentional control modulates a beta-centered motor set. The phenomenon is most prominent during the preparatory phase, when temporal expectations are formed [61].Taken together, these patterns suggest a shift from a fronto-parietal, alpha-dominant preparatory control state to a sensorimotor, beta-centric motor set across the transition from preparation to MI. This circuit-level account also explains our decoding results: phase-specific segmentation with late fusion leverages complementary information—attentional control during preparation and motor set during MI—thereby improving performance across window lengths. Potential clinical and assistive applications The PMS-MI paradigm elicits a brief, task-guided preparatory state immediately before MI. Because this state is ubiquitous in everyday actions and readily learned, the paradigm imposes a relatively low cognitive load; the task cue makes this state explicit and amplifies it in a controlled manner. In motor rehabilitation, this timing structure (preparation → MI) supports closed-loop protocols that deliver feedback or assistance (e.g., functional electrical stimulation or robot-assisted movement) at the moment motor intention is detected. Practically, the fixed preparatory cue and MI window also enable adaptive decoding: window length and decision thresholds can be titrated to training goals and patient progress—extending windows (e.g., 0–6 s) for sustained MI practice, or shortening them (e.g., 0–3 s) for rapid assistive device control—while leveraging preparatory-state features to increase the accuracy of motor-intent detection. These properties make the paradigm suitable for progressive, home-based schedules and for personalization across different clinical needs. Limitation and future work In this work, we investigated the preparatory movement state-based MI and decoded motor intentions. Our study included comparative analyses of MRCPs in the time domain, ERD in the frequency domain, and brain network patterns across different paradigms. We also explored decoding the preparation and MI phases using the MLED method. Despite these efforts, several directions remain for further exploration: In our experiments, the designed motor imagery focused on elbow movements. Future studies could explore more complex movement sequences involving multiple motion elements and compare them with actual movements to investigate their encoding patterns and decoding performance. Additionally, the MLED method utilized electrodes covering all brain regions and incorporated five frequency bands. Optimization techniques could be employed in future work to identify critical electrodes and frequency bands, reducing computation time and redundant features to enhance the system's practicality and efficiency. Although the present study reports offline results, in practice the 1-s cue is unnecessary, and the 2-s preparation can serve as a brief engagement/safety-gating stage (and can be shortened, e.g., to 1.5 s). Real-time deployment will be pursued via pre-computed GFT bases, sparse ROI connectivity with incremental updates, and lightweight linear classifiers, with pseudo-online and closed-loop evaluations planned. Finally, our cohort was modest (N = 14), which limits statistical power and generalizability. Future work will recruit a larger, demographically balanced cohort (potentially multi-site) with a priori power analysis to strengthen external validity, and will include patient cohorts to assess clinical transferability. Conclusions Building on prior research, this study integrates the preparatory movement state into a visually guided MI-BCI paradigm. Time-frequency analyses revealed that preparatory movement altered ERD oscillation patterns compared to the classical AO-MI paradigm. Brain topography showed distinct spatial patterns, while brain network analyses indicated reduced connection strength and fewer connections during MI under the PMS-MI paradigm. MI, as a complex process involving cognition, movement, and brain energy changes, informed the construction of a multilayered brain network matrix integrating intra- and cross-frequency features. Using GFT, we transformed network matrix energy into a complex domain and applied CSP to extract spatial differences for classification. Results demonstrated that optimal classification is achieved by segmenting and fusing features from both preparatory and MI phases. The primary contribution of this study lies in integrating the preparatory movement state into the MI paradigm. PMS-MI offers unique insights into neural resource allocation, with potential applications in clinical rehabilitation through individualized BCI training. This innovation enhances MI-BCI performance, while the proposed algorithm, adaptable to various scenarios by adjusting window lengths, improves system flexibility and practicality. The findings underscore the effectiveness of combining neuroscience and computer science to advance neuroengineering. By bridging theoretical innovation and practical applications, this work lays a robust foundation for future MI-BCI research, with implications for clinical therapies and beyond. Supplementary Information Supplementary material 1. Supplementary material 2. Supplementary material 1. Supplementary material 2.